Agent kept bending arms wrong, so this dev wired MuJoCo into an MCP server
usefulwrong · reddit · 2026-11-18
When simulating a push-up, the author's AI agent kept producing arm positions that failed the body model check. Instead of adding another "be careful" prompt, he connected MuJoCo and the MS-Human-700 body model via an MCP server.
Human Models MCP is open source and runs locally in Docker (github.com/usefulwrong/human-models-mcp). The agent can now:
- Discover body models, choose experiments, and run simulations
- Inspect joint motion, muscle activation, and contact forces
- Send computed movement to Blender for rendering
The demo uses a model with 700 muscle actuators; the palm push-up shown uses external joint assistance, with a separate muscle-controlled variant — neither validated against real motion capture.
He also raises an MCP interface design question: when simulation succeeds but the movement fails physical checks, should the tool return an error, or return the full result with failures and let the agent decide? He leans toward the latter but doesn't fully trust agents to read warnings.
Related event: Human Models MCP open-sources musculoskeletal simulation for AI agents(2 posts)→
More from coding & agent
- Doodlestein Shares His Multi-Agent Coding Stack: agent-mail + beads + ntm — doodlestein · 2026-11-18
- The bitter lesson is coming for software engineering: let agents discover abstractions — tylerjdunn · 2026-11-18
- MCP Agent Mail: open-source shared inboxes and coordination for multiple AI coding agents — doodlestein · 2026-11-18
- FrontierCS 2: benchmark makes AI agents beat published CS research blind — arena · 2026-11-18
- Keyframery adds Agent view on Vercel UI SDK to show what agents read — shadcn · 2026-11-18
- Microsoft Ignite demos Copilot at work as Chevron shows $295K monthly agent spend — donasarkar · 2026-11-18